L iQF d dl Z d dlmZ d dlmZ d dlmZmZmZm Z m
Z
d dl Z d dl m Z d dl
mZ d dlmZ dd lmZ dd
lmZ dd lmZ dd lmZ dd
lmZmZ ddl m Z e rd dl Z er>d dlm Z ddl m!Z! ddl"m#Z# ddl$m%Z% ddl&m'Z' ddl(m)Z) ddl*m+Z+ ddl,m-Z- ddl.m/Z/m0Z0 ddl1m2Z2 ejf e4 Z5 G d de Z6y) N)deepcopy)Path)
TYPE_CHECKINGAnyCallableOptionalUnion)nn)FullyShardedDataParallel)Dataset )GenerationConfig)is_deepspeed_zero3_enabled)is_fsdp_managed_module)Trainer)is_datasets_availablelogging)deprecate_kwarg)IterableDataset) DataCollator)FeatureExtractionMixin)BaseImageProcessor)PreTrainedModel)ProcessorMixin)PreTrainedTokenizerBase)TrainerCallback)EvalPredictionPredictionOutput)TrainingArgumentsc 8 e Zd Z edddd d'deedej f ded d
ed d eeed
df deeee e
ef f deed dee g df dee dee dge f dee d de
eej j eej j" j$ f dee ej& ej& gej& f f fd
Zedee
ef defd Z d(dee d ee e
d e
de e
ef f fd
Z d)d ed ee e
d e
dd!f fd"
Z d*dej d#e e
eej& ef f d$e d ee e
de
ee eej& eej& f f
fd%
Z d& Z xZ S )+Seq2SeqTrainer tokenizerprocessing_classz5.0.0T)new_nameversionraise_if_both_namesmodelr argsr
data_collatorr
train_datasetr zdatasets.Dataset eval_dataset)r r r r
model_initcompute_loss_funccompute_metricsr callbacksr
optimizers preprocess_logits_for_metricsc
t | ||||||||| |
| | | j j 7| j | j j }
|
| j
_ y y )N) r' r( r) r* r+ r# r, r- r. r/ r0 r1 )super__init__r( generation_configload_generation_configr' )selfr' r( r) r* r+ r# r, r- r. r/ r0 r1
gen_config __class__s b/mnt/ssd/data/python-lab/Trading/venv/lib/python3.12/site-packages/transformers/trainer_seq2seq.pyr4 zSeq2SeqTrainer.__init__6 sx $ ''% -! / + !*G
" 99 & & 2 44TYY5P5PQJ+5DJJ ( 3 gen_config_argreturnc t | t r t | }nqt | t r t | n| }d}|j r|j }|j }n|j rn| }t j || } |j d |S # t $ r }t t | dz d}~ww xY w)aW
Loads a `~generation.GenerationConfig` from the `Seq2SeqTrainingArguments.generation_config` arguments.
Args:
gen_config_arg (`str` or [`~generation.GenerationConfig]`):
`Seq2SeqTrainingArguments.generation_config` argument.
Returns:
A `~generation.GenerationConfig`.
NT)strictz'
Fix these issues to train your model.)
isinstancer r strr is_filenameparentis_dirfrom_pretrainedvalidate
ValueError)r< r8 pretrained_model_nameconfig_file_nameexcs r: r6 z%Seq2SeqTrainer.load_generation_config] s n&6 7!.1J =G~WZ<[D $8ao ! # %,,.#8#=#= (=(D(D%&--/ )7%)99:OQabJ U t , U SX(S ST T Us B# # C,CC ignore_keysmetric_key_prefixc |j }|j d @|j d /| j j | j j |d< |j d /| j j | j j |d< | j
j | _ || _ t | ) ||| S )a
Run evaluation and returns metrics.
The calling script will be responsible for providing a method to compute metrics, as they are task-dependent
(pass it to the init `compute_metrics` argument).
You can also subclass and override this method to inject custom behavior.
Args:
eval_dataset (`Dataset`, *optional*):
Pass a dataset if you wish to override `self.eval_dataset`. If it is an [`~datasets.Dataset`], columns
not accepted by the `model.forward()` method are automatically removed. It must implement the `__len__`
method.
ignore_keys (`list[str]`, *optional*):
A list of keys in the output of your model (if it is a dictionary) that should be ignored when
gathering predictions.
metric_key_prefix (`str`, *optional*, defaults to `"eval"`):
An optional prefix to be used as the metrics key prefix. For example the metrics "bleu" will be named
"eval_bleu" if the prefix is `"eval"` (default)
max_length (`int`, *optional*):
The maximum target length to use when predicting with the generate method.
num_beams (`int`, *optional*):
Number of beams for beam search that will be used when predicting with the generate method. 1 means no
beam search.
gen_kwargs:
Additional `generate` specific kwargs.
Returns:
A dictionary containing the evaluation loss and the potential metrics computed from the predictions. The
dictionary also contains the epoch number which comes from the training state.
max_lengthmax_new_tokens num_beamsrL rM ) copygetr( generation_max_lengthgeneration_num_beams acceleratorgathergather_function _gen_kwargsr3 evaluate)r7 r+ rL rM
gen_kwargsr9 s r: r[ zSeq2SeqTrainer.evaluate s N __&
NN< ( 0 /08 //;'+yy'F'FJ| $ >>+ & .4993Q3Q3]&*ii&D&DJ{ # # / / 6 6 %w +Yj kkr; test_datasetr c |j }|j d @|j d /| j j | j j |d< |j d /| j j | j j |d< | j
j | _ || _ t | ) ||| S )a&
Run prediction and returns predictions and potential metrics.
Depending on the dataset and your use case, your test dataset may contain labels. In that case, this method
will also return metrics, like in `evaluate()`.
Args:
test_dataset (`Dataset`):
Dataset to run the predictions on. If it is a [`~datasets.Dataset`], columns not accepted by the
`model.forward()` method are automatically removed. Has to implement the method `__len__`
ignore_keys (`list[str]`, *optional*):
A list of keys in the output of your model (if it is a dictionary) that should be ignored when
gathering predictions.
metric_key_prefix (`str`, *optional*, defaults to `"eval"`):
An optional prefix to be used as the metrics key prefix. For example the metrics "bleu" will be named
"eval_bleu" if the prefix is `"eval"` (default)
max_length (`int`, *optional*):
The maximum target length to use when predicting with the generate method.
num_beams (`int`, *optional*):
Number of beams for beam search that will be used when predicting with the generate method. 1 means no
beam search.
gen_kwargs:
Additional `generate` specific kwargs.
If your predictions or labels have different sequence lengths (for instance because you're doing dynamic
padding in a token classification task) the predictions will be padded (on the right) to allow for
concatenation into one array. The padding index is -100.
Returns: *NamedTuple* A namedtuple with the following keys:
- predictions (`np.ndarray`): The predictions on `test_dataset`.
- label_ids (`np.ndarray`, *optional*): The labels (if the dataset contained some).
- metrics (`dict[str, float]`, *optional*): The potential dictionary of metrics (if the dataset contained
labels).
rO rP rQ rR ) rS rT r( rU rV rW rX rY rZ r3 predict)r7 r] rL rM r\ r9 s r: r_ zSeq2SeqTrainer.predict s ^ __&
NN< ( 0 /08 //;'+yy'F'FJ| $ >>+ & .4993Q3Q3]&*ii&D&DJ{ # # / / 6 6 %w|Xijjr; inputsprediction_loss_onlyc 2 | j j r|rt |
|||| S d|v }| j | }t | dk( r&t
| d r| j j }d|v r|d |j d d|v r|d |j d t xs t | j }|j d| |d< |j }d|v rKd |v rG|d j |d j k( r(|j D
ci c]
\ } }
| d
vs | |
}} }
t! | j t" r t# j$ | j nt' j( } | 5 | j j* di ||} ddd | j j, j. rd | j j, _ | j j, }
j d |
j0 k r | j3 | |
j0 } nJ|
j4 >| j d |
j4 d
z k r | j3 | |
j4 d
z } t7 j8 5 |r| j; 5 |di |}ddd | j< 2| j= |d j? jA }n9t! tB r|d n|d j? jA }nd}ddd | j jD rddfS |r|d }|j d |
j0 k r | j3 ||
j0 }nM|
j4 A|j d |
j4 d
z k r"| j3 ||
j4 d
z }nd}| |fS c c}
} w # 1 sw Y HxY w# 1 sw Y IxY w# 1 sw Y xY w)a
Perform an evaluation step on `model` using `inputs`.
Subclass and override to inject custom behavior.
Args:
model (`nn.Module`):
The model to evaluate.
inputs (`dict[str, Union[torch.Tensor, Any]]`):
The inputs and targets of the model.
The dictionary will be unpacked before being fed to the model. Most models expect the targets under the
argument `labels`. Check your model's documentation for all accepted arguments.
prediction_loss_only (`bool`):
Whether or not to return the loss only.
gen_kwargs:
Additional `generate` specific kwargs.
Return:
tuple[Optional[float], Optional[torch.Tensor], Optional[torch.Tensor]]: A tuple with the loss, logits and
labels (each being optional).
)ra rL labelsr rZ rQ NrO synced_gpusdecoder_input_ids)re decoder_attention_maskFr
loss )#r( predict_with_generater3 prediction_step_prepare_inputslenhasattrrZ rS popr r r' rT shapeitemsr@ r summon_full_params
contextlib nullcontextgenerater5 _from_model_configrO _pad_tensors_to_max_lenrP torchno_grad compute_loss_context_managerlabel_smootherdetachmeandictra )r7 r' r` ra rL r\
has_labelsdefault_synced_gpusgeneration_inputskvsummon_full_params_contextgenerated_tokensr8 outputsrh rc r9 s r: rk z Seq2SeqTrainer.prediction_step s > yy..2F7*v4HVa +
'
%%f- z?a GD-$@ ))..0J * $K)@)H NN; ' : %*\*B*J NN< ( 8 : `>TUYU_U_>`$.NN=BU$V
=! "KKM
) )#'88!(+115FGZ5[5a5aa "( ! A1?
% 7 7
C'') # ( V 2tzz 2 2 U5F U* U V :: ' ' : :>CDJJ ( ( ; ZZ11
! !" %
(=(= = # ; ;S>ST**66<<;KjNgNgjkNk;k 55fj>W>WZ[>[\F%v--e! V V&. . s= * O-7O- O3 P
0 P 9A?P
3O= P
P
Pc | j Yt | j d rC| j j | j j n| j j }nL| j j
j !| j j
j }n t
d |t j |j d |f|j |j z }||d d d |j d f< |S )N pad_token_idzSPad_token_id must be set in the configuration of the model, in order to pad tensorsr )dtypedevicerg ) r# rn r eos_token_idr' configrH rx onesrp r r )r7 tensorrO r
padded_tensors r: rw z&Seq2SeqTrainer._pad_tensors_to_max_lenp s ,9N9NP^1_ ((55A %%22**77
zz --9 #zz 0 0 = = !vww$uzz
\\!_j ) fmm(
06
a+6<<+++, r; ) NNNNNNNNNN)NNN)NNeval)Ntest)N) __name__
__module__ __qualname__r r r r
Moduler r~ rA r listtuplerx optim Optimizer lr_schedulerLambdaLRTensorr4 staticmethodr r6 floatr[ r_ r boolrk rw
__classcell__)r9 s @r: r! r! 5 s [+=wdhi @D.226Y]EI
@D04HL7;jvhl $6 / :;<$6 *+$6 / $6
g/@BT&T UV $6 uWd3<.@%@ AB
$6 # m n
$6 Xb*;&; <=$6 $H-$6 "(,<+=t+C"DE$6 D!2 34$6 (5;;#8#898EKKD\D\DeDe;ffg $6 (0%,, 9UW\WcWc9c0d'e $6 j$6L )uS:J5J/K )P` ) )Z +/+/!' 6l w'6l d3i(6l 6l
c5j
6lv ,0!' >k >k d3i(>k >k
>kJ ,0 m.yym. S% c 1223m. # m.
d3i( m.
x 6 8NN Om.^ r; r! )7rs rS r pathlibr typingr r r r r rx r
torch.distributed.fsdpr torch.utils.datar generation.configuration_utilsr integrations.deepspeedr integrations.fsdpr trainerr utilsr r utils.deprecationr datasetsr data.data_collatorr feature_extraction_utilsr image_processing_utilsr modeling_utilsr processing_utilsr tokenization_utils_baser trainer_callbackr
trainer_utilsr r
training_argsr
get_loggerr loggerr! ri r; r: r s @ @ ; $ < > 5 1 . 00@:/0@1?0
H %M W M r;